Transformer-based relation extraction and concept normalization using an annotated clinical trials corpus

Abstract

Healthcare professionals manually review electronic health records to select patients who meet eligibility criteria of clinical trials. Natural language processing offers a complement for this task, although few initiatives exist in Spanish. We present version 3 of the CT-EBM-SP corpus of 1200 clinical trials (292173 tokens), annotated with 23 entity types and 18 relation types, covering Unified Medical Language System (UMLS) semantic groups, drug-related information, temporal data, and negation/speculation. We encoded 11 attributes (e.g., event temporality and experiencer status) and normalized entities to UMLS Concept Unique Identifiers. The corpus contains 87037 entities, including nested and discontinuous entities, 16597 attributes and 68206 relationships. Inter-annotator agreement (IAA) achieved average F1 values of 0.861 (entities), 0.810 (attributes), and 0.791 (relations). 81.75% of entities were normalized (IAA: F1 = 0.966). We benchmarked this dataset by fine-tuning Transformer models for relation extraction (RE) and medical concept normalization (MCN). In RE, the average F1 ranged from 0.858 to 0.879, and for MCN, the accuracy at rank 1 was 0.896. The corpus and models are publicly available.

Ähnliche Arbeiten